What Donald Trump is Tweeting (Analyzing Tweets with NLTK and Pandas)

How does @realDonaldTrump (Donald J. Trump) tweet? Before the President-elect became the president-elect, I didn’t pay much attention to his tweets, but I did know that he seemed to have a unique style of writing them. To me, it felt like he tweeted like he spoke in public. But what was that tone or Trump brand?

From a quick glance of his account, the usage of exclamation points seemed prominent. I wondered if that was consistent, and I wondered what else I could find. So I ran an analysis Trump’s tweets as well as @HillaryClinton, @CNN and @FoxNews’ for some comparison.

My strategy for selecting the accounts:

  1. Find another individual user’s account with opposing views, but who had some similar goals over the last year + (to be come President of the United States)
  2. Compare the individual users’ accounts with “objective” news sources’ Twitter accounts. Since source objectivity is always a topic of hot debate (or a hot topic of debate), I took @CNN and @FoxNews.

The Method

  • Pull as many of Trump’s Tweets as I could via Twitter API. The best service to perform this operation is /user_timeline. You can only get the last 3,200 tweets from any particular user handle, which is unfortunate because plenty of accounts have authored far more than 3,200 tweets (e.g. Trump has written over 30,000 tweets).
  • Take the tweets and do some Natural Language Processing with Python’s NLTK.
    • Get the parts of speech tags for every word in every tweet
    • Do word frequency counts
    • Classify the sentiment of the tweets
  • Classify the reading level or difficulty of each word. Determine the reading level of the account. 
    • I didn’t implement this. I started, but then I got distracted.
  • Pull out an aggregate view of some other interesting tweet data, like #HASHTAG usage.
  • Compare data.

And Now, The Data

Punctuation! Punctuation! Punctuation! Trumpunctuation?

If you’ve seen Trump speak, you probably may have noticed his strong intonations and general emphatic demeanor. His tweets seem to capture this partially through punctuation alone. His words are sprinkled with the strongest punctuation mark in the English language, the exclamation mark.

“!” occurs 2336 times over 3200 Trump tweets. 

And “!” occurs at least once 1954 times over 3200 Trump tweets.

Which means that 61% of Trump’s tweets contain an exclamation mark (based on my 3200 tweet dataset). This seemed astonishingly high; when I first pulled this I thought my code was incorrect. So I opened my file of tweets and eyeballed it to confirm. I figured if 60+ % of tweets contained an exclamation mark, then it would be easy to confirm my sanity (or lack thereof) from the text file.

trump's usage of exclamation mark

Some of Trump’s Tweets

The highest frequency of “!” in a single tweet was five. And that occurred in the following tweet:

“#WheresHillary? Sleeping!!!!!”
Which was Retweeted 27,158 times and Favorited 61,084 times. Created on August 20, 2016.


Here’s the !!!! comparison among the entire group:

@realDonaldTrump Absolute Count (!) @HillaryClinton Absolute Count (!) @CNN Absolute Count (!) @FoxNews Absolute Count (!)
2336 171 43 134

Adjectives Are Very Great

Trump’s top twenty favorite adjectives are listed below, along with the top 20 adjectives from the comparison group. It’s not really a surprise that “great” was number one for Trump because, among other things, his campaign’s slogan was “Make America Great Again.” Notably, I removed stopwords and only grabbed lowercase words to help filter out some noise from the data. Relative frequency is the percentage relative to the other adjectives in the dataset.

trump_word abs_frequency rel_frequency
great 211 4.953052
bad 81 1.901408
many 81 1.901408
big 77 1.807512
last 66 1.549296
new 56 1.314554
good 50 1.173709
much 34 0.798122
amazing 32 0.751174
total 30 0.704225
wonderful 30 0.704225
nice 27 0.633803
massive 25 0.586854
first 25 0.586854
presidential 23 0.539906


Below is the list of top 20 adjectives from @realDonaldTrump, @HillaryClinton, @CNN, @FoxNews from their last 3200 tweets.



General Sentiment of Tweets

To calculate the sentiment of the tweets, I used a function in the nltk library called demo_liu_hu_lexicon() which classifies each word of the sentence as Positive, Negative, or Neutral, and then does a basic count of each word-classification category. Whichever group has the highest count is how the text will get assigned. There are definitely better ways to do this. I considered an integration with IBM’s Watson, but time was doing its thing, being time, and being of the essence and such.

@realDonaldTrump @HillaryClinton @CNN @FoxNews
Positive 0.463750 0.468750 0.304688 0.235625
Neutral 0.285313 0.401875 0.425938 0.455937
Negative 0.250937 0.129375 0.269375 0.308437

Top Hashtags

This section contains a top 5 hashtag summary table for the entire analysis group, and also has the top 20 hashtags for each account listed afterward. You can infer what you will from this data.

I did find it interesting that the top-used hashtag by Trump was one of self-promotion, and Hillary Clinton used a lot of hashtags relating to the debates. Just looking at the hashtag data makes me think that Trump’s social media strategy was much stronger throughout the campaign.

In addition, all four Twitter accounts had at least one hashtag with “Trump” in it. From a marketing perspective, that’s good brand awareness.

Top 5 Hashtag Summaries

@realDonaldTrump Top 5 Hashtags @HillaryClinton Top 5 Hashtags @CNN Top 5 Hashtags @FoxNews Top 5 Hashtags
#Trump2016, #MakeAmericaGreatAgain, #MAGA, #AmericaFirst, #DrainTheSwamp #DebateNight, #DemsInPhilly, #VPDebate, #RNCinCLE, #debate #CNNHeroes, #CNNSOTU, #Aleppo, #CNNNYE, #JadonAndAnias #KellyFile, #Trump, #FoxNews2017, #Hannity, #Christmas

@realDonaldTrump Hashtags

abs_frequency hashtag rel_frequency_pct
238 Trump2016 15.130324
202 MakeAmericaGreatAgain 12.841704
111 MAGA 7.056580
79 AmericaFirst 5.022250
78 DrainTheSwamp 4.958678
57 ImWithYou 3.623649
57 BigLeagueTruth 3.623649
53 VoteTrump 3.369358
38 CrookedHillary 2.415766
36 Debate 2.288620
35 TrumpTrain 2.225048
34 TrumpPence16 2.161475
24 Debates2016 1.525747
22 ICYMI 1.398601
20 SuperTuesday 1.271456
18 VPDebate 1.144310
15 RNCinCLE 0.953592
14 Debates 0.890019
13 WIPrimary 0.826446
12 ThankYouTour2016 0.762873

Below are the top hashtags for the entire group.

It’s also worth sharing how many total hashtags were used in the last 3200 tweets of each Twitter account in the dataset group:

@realDonaldTrump @HillaryClinton @CNN @FoxNews
total_hashtags_used 1573 482 143 810


Top Favorited Tweets

Top Favorited of @realDonaldTrump

text favorite_count created_at
Such a beautiful and important evening! The forgotten man and woman will never be forgotten again. We will all come together as never before 639428 Wed Nov 09 11:36:58 +0000 2016
TODAY WE MAKE AMERICA GREAT AGAIN! 577008 Tue Nov 08 11:43:14 +0000 2016
How long did it take your staff of 823 people to think that up–and where are your 33,000 emails that you deleted? 296236 Thu Jun 09 20:40:32 +0000 2016
The media is spending more time doing a forensic analysis of Melania\’s speech than the FBI spent on Hillary\’s emails. 248370 Wed Jul 20 15:36:06 +0000 2016
Just had a very open and successful presidential election. Now professional protesters, incited by the media, are protesting. Very unfair! 234619 Fri Nov 11 02:19:44 +0000 2016
Love the fact that the small groups of protesters last night have passion for our great country. We will all come together and be proud! 224497 Fri Nov 11 11:14:20 +0000 2016
Nobody should be allowed to burn the American flag – if they do, there must be consequences – perhaps loss of citizenship or year in jail! 216064 Tue Nov 29 11:55:13 +0000 2016
Fidel Castro is dead! 212487 Sat Nov 26 13:08:11 +0000 2016
This will prove to be a great time in the lives of ALL Americans. We will unite and we will win, win, win! 204239 Sat Nov 12 15:05:33 +0000 2016
A fantastic day in D.C. Met with President Obama for first time. Really good meeting, great chemistry. Melania liked Mrs. O a lot! 194927 Fri Nov 11 02:10:46 +0000 2016


Top Favorited of @HillaryClinton

text favorite_count created_at
“I never said that.” \u2014Donald Trump, who said that. #debatenight 160098 Tue Sep 27 01:19:47 +0000 2016
Where was this kind of comedy last night? 134956 Fri Oct 21 17:51:34 +0000 2016
“Trump just criticized me for preparing for this debate. You know what else I prepared for? Being president.” #DebateNight 112647 Tue Sep 27 02:01:59 +0000 2016
Women have the power to stop Trump.\\n\\n\\n 111546 Fri Oct 07 23:54:35 +0000 2016
“Nobody respects women more than me.” \u2014Donald Trump earlier tonight\\n\\n”Such a nasty woman.” \u2014Donald Trump just now #DebateNight 108114 Thu Oct 20 02:36:56 +0000 2016
Don\’t stand still. Vote today: #ElectionDay #MannequinChallenge 95882 Tue Nov 08 11:47:18 +0000 2016
Happy birthday to this future president. 94281 Wed Oct 26 13:03:18 +0000 2016
RT if you\u2019re proud of Hillary tonight. #DebateNight #SheWon 93121 Thu Oct 20 02:37:43 +0000 2016
RT this if you\’re proud to be standing with Hillary tonight. #debatenight 92721 Tue Sep 27 02:45:44 +0000 2016
This is horrific. We cannot allow this man to become president. 91471 Fri Oct 07 20:55:13 +0000 2016


Top Favorited of @CNN

text favorite_count created_at
At least 60 people have been hurt in an explosion at a fireworks market near Mexico City, local media report.\u2026 22570 Tue Dec 20 22:49:24 +0000 2016
This little boy is the newest face of OshKosh B\’gosh\’s holiday ads after initially being turned down by a talent ag\u2026 22506 Sat Dec 10 07:01:29 +0000 2016
“We\’ve been friends for a long time”: Kanye West and President-elect Trump appear together at Trump Tower\u2026 19082 Tue Dec 13 14:59:55 +0000 2016
This little boy is the newest face of OshKosh B\’gosh\’s holiday ads after initially being turned down by a talent ag\u2026 12099 Sat Dec 10 23:45:25 +0000 2016
Clinton jokes at Reid portrait unveiling: “After a few weeks of taking selfies in the woods, I thought it would be\u2026 11868 Thu Dec 08 22:23:56 +0000 2016
BREAKING: President Obama vows retaliatory action against Russia for its meddling in the US presidential election\u2026 9767 Fri Dec 16 01:36:31 +0000 2016
More Americans voted for Hillary Clinton than any other losing presidential candidate in US history 9273 Wed Dec 21 23:28:00 +0000 2016
This little boy is the newest face of OshKosh B\’gosh\’s holiday ads after initially being turned down by a talent ag\u2026 9117 Mon Dec 12 00:42:57 +0000 2016
Mariah Carey, Adele, Elton John and Lady Gaga bring the holiday spirit in special Christmas-themed \’Carpool Karaoke\u2026 7475 Mon Dec 19 05:30:24 +0000 2016
\’Dear world, why are you silent?\’: Desperate pleas from inside Aleppo 7355 Wed Dec 14 19:00:04 +0000 2016


Top Favorited of @FoxNews

text favorite_count created_at
JUST IN: President-elect #Trump announces @Sprint will bring 5,000 jobs back to the U.S., and OneWeb will hire 3,00\u2026 18974 Wed Dec 28 22:17:35 +0000 2016
.@JudgeJeanine: “Michelle, you may not realize it, but Americans rejected you and everything you stand for.”\u2026 15054 Sun Dec 18 17:17:51 +0000 2016
#Breaking News: President-elect @realDonaldTrump has garnered the 270 #ElectoralCollege votes needed to become pres\u2026 11164 Mon Dec 19 22:35:53 +0000 2016
Peters: “Without the least exaggeration, we can say that President Obama has been the worst foreign policy presiden\u2026 10263 Sat Dec 31 03:32:46 +0000 2016
.@realDonaldTrump: “Michelle Obama said yesterday that there\’s no hope, but I assume she was talking about the past\u2026 10021 Sat Dec 17 22:43:31 +0000 2016
.@realDonaldTrump: “We have to protect Israel. Israel, to me, is very, very important. We have to protect Israel.” 10021 Sun Jan 01 03:26:17 +0000 2017
Giuliani: \u201cThe U.S. Constitution doesn\u2019t give anyone in this world the right to come to the U.S. That\u2019s a privilege\u2026 8885 Wed Dec 21 03:41:24 +0000 2016
.@KatrinaPierson: This president is the divider-in-chief. His entire political career revolved around racism, sexis\u2026 8700 Fri Dec 30 01:43:04 +0000 2016
.@GovMikeHuckabee: Can you name me one Muslim country that welcomes Christians to build & protect churches? No, you\u2026 8419 Thu Dec 29 01:39:42 +0000 2016
DJT: “I think the Democrats are putting it out because they suffered 1 of the greatest defeats in the history of po\u2026 8175 Sun Dec 11 19:08:47 +0000 2016

About the Data

If you would like to access some of the code/data, it is publicly available on my GitHub repo. I’ve also included all four files that contain all of the tweets on which I ran the analysis in the data/ directory.

I had plans to include Date/Time data analysis in this post and many other things (if you’d like to see more data, let me know), but you have to stop somewhere great!!!!!! #MakeCodingProjectsSmallAgain

Coding, How-To, New York City, Random

How a Copywriter Becomes a Software Engineer

When I graduated college, I had a degree in Philosophy and a desire to work as a writer — specifically as a journalist.

“So much liberal, so much arts.”

This eventually became, “so much for liberal arts!”

And for a while, writing was precisely the work that I did. I moved to NYC and started my career as a creative (copywriter and strategist) at a digital ad agency. I was given a cheap laptop and a seat without much leg room and was put right to work — thank God the work didn’t involve me moving because my legs were asleep from the lack of space (not true, unless I was in outer space, or something).

But before I started that job as a copywriter, I had this crazy tech company idea: Newspanion — “it’s like Pandora, but for news.” Back in 2010/2011, there weren’t a lot of Pandoras for news, so it was a novel idea, which is no longer the case. My business partner, Andrew, was a guy I met some many months earlier in line for a bar (apparently my early-20s self thought it was acceptable to wait in lines to access alcohol).

Together, he and I wanted to start a tech company…

and we had pretty much zero tech skills.

What does that mean? Well, we weren’t going to be technologically engineering anything, unless you count engineering e-mails to try to find a technical cofounder.

So here I was, a copywriter/strategist moonlighting as a tech entrepreneur — writing copy by day and business plans by night. Collectively, Andrew and did a lot of research, wrote a lot of business stuff, sent a lot of business e-mails, and tried our damned-est to find a technical cofounder.

But it never worked out; something always went wrong.

It was around my 5th or 6th month at the agency when I decided that I would try to build a prototype for our startup so we could pitch investors. I didn’t know where to start.

I hadn’t written any HTML since Geocities in 1997.

Hell, I didn’t even know if HTML would do it. I did some research, and everyone was always talking about this Ruby on Rails thing. That sounded like it would do the trick because it sounded fancy enough. But some people said I should learn Ruby before Rails (and I was like, “Where’s the train platform again? Is that near the jewel excavation site?”).

I took their advice and bought a book on Ruby. And the next thing I noticed was that everyone using Rails seemed to be using Macs. In fact, lots of software engineers were on Macs, and I had a rickety-ass, 7-year-old Toshiba that looked a lot like something you could use as an ad-hoc baseball plate (if you were actually caught in a base-less-baseball-game pinch). This laptop looked like a stove, or an ironing board, or a model UFO. Actually, not a model UFO, that would look cool; this thing was so uncool.

toshiba laptop

My Toshiba laptop and netbook. Remember netbooks?!

Coming to terms with these facts, I bought a 13-inch Macbook pro, which I’ve been using as a reliable (and expensive) paperweight recently. When I unpackaged that puppy though, it was like coding Christmas. Ruby was already installed! Things were going to be so great!

… And pretty much every day for a few months, I went to work, wrote some ads and copy.

Then I went home and wrote some really, really bad code.

Actually, a lot of the time I didn’t even write it, I copy-pasta’d it. Those early days were filled with plenty of soup and copy pasta. I might as well have called myself a sous-do chef!

I really had no idea what I was doing. I didn’t have a good framework for understanding the material. I found my studies in philosophy to help me during these Boeing-747-turbulent times by helping me ask good questions (I also had plenty of others… caution: article contains bad words!):

  • “What is the problem I am trying to solve here?”
  • “What is this doing? I want to understand this.”
  • “If this works and that works, what happens when I remove this? Does it still work?

Good ol’ deductive reasoning. I never thought I’d be deducing bug origins (of type software or creepy-crawly) when I was sitting reading Nichomachean Ethics, but gosh was I wrong.

I just kept plugging away, until one day,

I had a really shitty-spaghetti prototype (Bon Appetite!)

of our web application.

But hey, it worked, and now we could at least show people our idea. Long story short: we showed and showed, but just never could get the traction we were looking for, and the startup died.

Meanwhile though, back in real-life world, I was still a copywriter and strategist, writing ads that you probably didn’t even notice while surfing the interwebs. About 12 months into that job, I decided I’d try a hackathon. At the event, I luckily sat next to an engineering wizard, Toby, (basically the Harry F’ing Potter of coding), and we won. With the taste of victory so fresh, I went to another hackathon and won that one too. I will always hold that we won because of our teams; my teammates defined excellence. At these hackathons I provided product/creative contributions, in addition to writing mangled HTML/CSS/Javascript.

It was about this time that I was feeling pretty awesome about my new skillz. I mean, I could actually pay the billz now!

Fast forward a little bit to 2012. I quit my job as a copywriter and joined AppNexus, that global ad tech company you may have heard of. I started at AppNexus as a technical account manager (TAM), since there was no way I could really be hired as a software engineer (well, maybe, if I only did front-end). But still, at tech companies, there are lots of people more technical than you when you are a self-taught web developer. And so I was a TAM, TAMing away at the beat of my drum. TAMing was great because it taught me that I had enormous gaps in my computin’ knowledge. I didn’t know how to use CURL or write SQL statements when I started at AppNexus. But most important of all, I learned that I didn’t really know how to figure things out on my own yet — I mean the type of figuring out that doesn’t include Stack Overflow, but instead, includes RTFMing.

Just RTFM!!

Just RTFM!!

As a TAM, I found myself doing all sorts of repeated tasks in Excel, which I didn’t like much, so I took up scripting and Python.

I’ll tell you what, my first Python scripts were so bad

that someone probably should have written a script…

to rewrite all of mine.

Sometimes I wondered if I should have stuck to writing movie scripts.

With time, though, I got faster and better, and was on to the next challenge: Data Analysis with Python. I moved into a new role as an analyst. Being a data analyst is cool because you get to decide on which tools you want to use to conduct your analysis, and then you also get to decide how you want to display the data. I chose Python for everything I did, just to get better. And better I got. (I also would often do things the hard way just to get the most learning out of the task.)

In August of 2015 I moved into a Software Engineering role,

approximately 1540 days after my initial efforts

to learn how to build software back in May, 2011.

Granted, I took plenty of breaks in between and probably could have accelerated the process by years by going for an internship, or more school (… ew). But I didn’t go that route, because it’s about the journey, not how you get there! Wait a minute. It’s how you got there, not the journey. Uh, scrap the platitudes… Did someone say journey?

What I meant to say is: Don’t stop believing.


How to Retrieve and Analyze Your iOS Messages with Python, Pandas and NLTK

I’m one of those people that keeps every text message I send or receive — I never delete them. Meet a girl at a bar, text her the next day and never hear back from her? I keep that. Weird wrong-number texts? I keep those too. Ex-girlfriend texts? Definitely keepers.

I had 65,378 messages on my phone at the time of writing this post.

I’m not a digital hoarder or anything, but I primarily do this because I like the idea of being able to search through the past. But, digital hoarder or not, collecting anything takes up some sort of space, and when I found that my text messages were taking up 4GBs of space on my phone, I decided it was time to back them up. It was at that point that I realized I could also probably analyze them.

As it turns out, you can do this, and I’ll tell you how. For this project, I used Python/Pandas/NLTK for the analysis and an iPython Notebook to render the datasets. I’ve also uploaded the code to GitHub, which you can view here.

An overview of the steps to make this happen:

  1. Sync/back up your iPhone because the messages need to be stored on your computer.
  2. Load the SQLite file and retrieve all messages
    • You can follow the directions for retrieving the right file here.
  3. Analyze those mensajes (I used Pandas)!!

Let’s get into some details.

You need to sync and back up your phone’s contents to your computer. There’s a great post on how to do this here. In case you want to skip that read, you’re ultimately getting a file with the text messages in it; copying it and moving it into your working directory.

You can find the file with this bash command:

$ find / -name 3d0d7e5fb2ce288813306e4d4636395e047a3d28

Now, loading the SQLite file — you can actually see what’s in this file via the command line:

 $ sqlite3 3d0d7e5fb2ce288813306e4d4636395e047a3d28 

Then you can check out the available tables:

sqlite> .tables
_SqliteDatabaseProperties chat_message_join
attachment handle
chat message
chat_handle_join message_attachment_join

From here, the main tables I found useful were “message” and “handle.” The former contains all of your text messages, and the latter contains all of the senders/recipients. I only wrote code around the messages table, primarily because I could never figure out how to make a join between message and handle, but that was probably something trivial that I overlooked. Please tell me how you did it, if you did!

Continuing on, the message table has lots of columns in it, and I chose to select from the following:

['guid', 'service', 'text', 'date', 'date_delivered', 
'handle_id', 'type', 'is_read','is_sent', 'is_delivered',
'item_type', 'group_title']

The key field is “text,” which is where the content of the message is stored, which includes emojis! (A cool thing is that your emojis will show up if you try to plot them in something like an iPython notebook. You could run an entire analysis on emoji usage…)

My analysis, however, ultimately breaks down into two pieces:

  1. Analyzing the content of the “text” field (excluding emojis).
  2. Analyzing the messages themselves (for example, total text messages, or, what I sent vs. what I received, for instance).

For #1, I wrote code that:

  • Classifies all words and assigns a part of speech to them, then check the counts of each part of speech.
    • You should get a table looking like this.

      You should get a table looking like this.

  • Counts the number of times each word appears in the dataset, and gives an overview of the dataset:
    • total_words_filtered
  • Excludes boring words, like prepositions, and words that are < 2 characters.
  • Classifies all words as is_bad=1 or 0. I did this by using a .txt file full of bad words, found here:
  • Plots usage of bad words
    • I’d love to show you my plot, but let’s just assume I never swear…

For #2, the code allows you to:

  • Plot the number of text messages received each day (check out the spike on your birthday or during holidays). You can see my data below has a huge gap (that’s when my phone was replaced and not backed up for many months. My timestamp conversions are also apparently incorrect, but I haven’t looked into it.
    • The timestamp conversion is off, so someone can fix that... we're not in 2016, yet... Are we??

      The timestamp conversion is off, so someone can fix that… we’re not in 2016, yet…Or am I??

  • Count the number of sent versus received messages.

Anyway, I hope you can get some use out of this, and instead of blabbing on about the code here, I’ll just let you read it and use it on your own. Please check out my git repo, and please reach out to me with questions, comments, etc.

Coding, Hackathon, New York City

Coming Through In the Clutch at Sports Illustrated’s Hackathon

“Hackathons are, like, my new favorite thing,” said not a lot of people. But I’m saying it, because I like building proof of concepts in a short period of time.

Hackathons are great reminders that great things can be built in no time at all.

Earlier this month, I participated in Sports Illustrated’s first hackathon, and this is a brief recounting of that sugar-filled, sleep-deprived, product-building experience. (No sports were actually played, unfortunately.)

We showed up around 9:30 a.m. because we were told that more people RSVP’d than the event space could handle. When I arrived, there were already about 10 over-achievers who had gotten there before me. I wanted to say, “damn them,” but then I realized, we were all probably sewn from similar fabric — the kind that welcomes waiting in lines to do work on a Saturday morning for the prospect of a unknown prizes (prizes had not been announced yet).

Around 10 a.m., the event started and we were allowed into the space (Time Inc. Photo Studio), which had a nice view.

The proof is in the pudding.

The proof is in the poster pudding.

A good view for building things that matter.

A good view for building things.

I can't say I read those two issues.

I can’t say I read those two issues.

There was coffee, donuts, bananas and plenty of other snacks to get fired up. My teammate, Brett, and I picked out a table and got settled in. Within about 15 minutes, there were speakers giving presentations to the group of hackers. This was a bit odd, and unfortunately, fell on the note of contrived. But alas, no donut is free. And the final speaker, Alex Bresler, said something that inspired what we would actually be building.

“It’s really difficult to determine who is more clutch than someone else.”

When I heard this, I thought, “Why? We can do that. We’ll do that.” I messaged Brett (sitting next to me), and asked what the thought about building something that rates the “clutchness” of players in sports, and allows users to compare different players. He responded that he liked that idea.

Left: Brett paying attention to speakers. Right: Me not paying attention.

Left: Brett paying attention to the presentations. Right: Me not paying attention.

Shortly thereafter, the speech wrapped up, and one of the guys sitting at our table asked us what we were working on. I stated that we had an idea, but we weren’t totally sure. We asked him what he did, and he said he worked on “statistical modeling,” or something like that. I was skeptical, as I’m sure Brett was as well — I could see it on his face. I decided to give a little more info about what we were working on, and this statistician dude asked if he could join our team. I thought, “I don’t know anything about this guy, but, why the hell not. He could be great.” I said, “yeah sure, but let me ask Brett.” Brett was not yet convinced, but finally, he decided to say yes too. And so we accepted this stranger onto our team.

I asked him what his e-mail was so I could invite him to our git repo, and he said, [actual-email-address]

I laughed a good laugh inside. I knew regardless of what would happen, I was probably going to learn something. Our new teammate’s name was Dan (and still is, actually).

After officially becoming team-complete, I gave Dan the idea straight up, he seemed to like it. We discussed how we would build it, and then we were off and running. The unfortunate time was roughly 1 p.m.; this was not a true 24-hour hackathon.

We coded through the “early” night and Dan worked up some awesome model for calculating clutchness. By around 1 a.m., we had something pretty serviceable. Subsequently, this was about the same time my bed started to call me. The problem was, my bed was miles away, and it started to rain, and I did not want to walk anywhere in the rain. We also had some nasty bugs in our code that needed to be squashed, and squash them we did. It was around this time, maybe a little bit earlier, that we started to pair program, because this is when 1 brain is worth about 1/2 a brain. Dan took off for his bed sometime around 2 a.m., I think? Not sure.

Around 4 a.m., delirium struck, and I told Brett I was going to need some nappy time. Brett, as though he were cast of iron, turned to me and said that he was probably going to keep working. I admired that, but I also admired my potential time with the couch. I briefly napped and I returned to our table 30 minutes later, hardly refreshed, but it was better than nothing.

Brett was working on some insane bash command that was downloading videos, compressing them, converting them to another format, and probably sending them to the International Space Station. I kinda thought it was magical, but that’s because everything was magical at that hour, including water. A toothbrush would have been magical, too.

The sun came up; that was good. Dan came back early that morning, and pushed all of his work to our repository. Suddenly, we not only had data, we had a real mathematical model computing the clutchness, or, Clutch Rating of a particular player. We were ecstatic.

As we got closer to stopping-time, I built out a brief outline of our pitch, and we made some last-minute tweaks. The judges arrived (including the CMO of Sports Illustrated, Damian Slattery; the Executive Editor of Sports Illustrated, Ryan Hunt; and the CTO of Time Inc, Colin Bodell) and said some short words, and with that, the pitches started.

Check out that high-tech digital display.

Check out that high-tech digital display of the pitching order.

We got up on stage, and it felt like we nailed it. Our product was quite complete, by hackathon standards. Users could:

  • Search through players, see a snapshot of each player and their overall Clutch Rating.
  • Click on a player and see more stats about the player, including a profile photo.
  • View all of a particular player’s plays that went into computing their overall clutch rating.
  • Click on a play to watch game footage of that play (the product of Brett’s insane shell command).

We were the last team to pitch, and after finishing, the judges went on with their judging.

About 5 minutes later, the winners were announced, and we took first place.

This is what winning looks like.

This is what first place looks like. From left to right: Alex (me), Dan, Colin Bodell, Brett

Plenty of post-hackathon networking ensued, and we went to get a few whiskeys — not that we needed them. We were already mostly drunk with delirium. After the second or third drink, I turned into what some would call, a zombie. I promptly made my way to the train station without eating anyone, and found my bed.

The next day, we were all off to work like nothing ever happened.

Check out our project at  

Coding, Hackathon, New York City

Winning at TechCrunch Disrupt’s Hackathon — 1st Place for Microsoft Outlook Hack

On Friday, May 1st, I was moving to a new apartment with a big ol’ U-Haul truck. On Saturday, May 2nd, I was at Techcrunch Disrupt for the hackathon, aiming to win. And win we did.

Disrupt Logo

Earlier that morning, Jeff and I discussed some potential ideas over coffee, omelettes, and flapjacks. We came down to one: create a timesheet automatically by scanning your sent e-mails and evaluating your past meeting invites. Some people know how infuriating it can be to create a timesheet for clients, because if you don’t log your time immediately, you can forget what you spent the time doing in the first place. So, this was a hack to solve that problem.

One line I was sort of OK to wait in.


When we started, I was already physically exhausted, but my brain was still fresh, and I was ready to rip.  Because we were a small team of three, we knew we had to make something with a reasonable scope.

Max productivity!

Max productivity!

We hacked and hacked, as those who hack are wont to do. And this is the rest of the story in pictures:

Backstage before the pitch!

Backstage before the pitch! 

The pitch video!

Making those bills.

Making those bills.

And this was my reward!

And this was my reward!

Jeff put together a nice write-up of the win here.

And that’s a wrap.

Coding, How-To

How to Create Geo HeatMaps with Pandas Dataframes and Google Maps JavaScript API V3

Get excited because we’re going to make a heatmap with Python Pandas and Google Maps JavaScript API V3. I’m assuming the audience has plenty of previous knowledge in Python, Pandas, and some HTML/CSS/JavaScript. Let’s begin with the DataFrame.

The DataFrame

First, you’re going to need a dataframe of “addresses” (can be a physical address, or even just a country name, like USA) that you eventually want to plot. (For the sake of simplicity, I’ll try to refer to the “address” as the “geo” for the rest of this document.) Second, since you are planning on using a heatmap, you’re going to want some sort of number that represents the weighted value of that row in comparison to other rows.

Let’s say your DataFrame looked like this:

grouped_country_df = main_df.groupby('country')\
                            .agg({'pink_kitten': lambda x: len(x.unique())})\
                            .sort('pink_kitten', ascending=False)
print grouped_country_df
geo_name count_of_pink_kittens
USA 3430
Spain 577
United Kingdom 352
Israel 292
Austria 196
Argentina 151
India 133
Singapore 66

Now you have a list of geos and some values to use as the weight when later creating the heatmap. But to plot these points, you’re going to need some lat and long coordinates.

Getting Lat Long Coordinates from Google Maps API

If you have a list of geos or “addresses,” you can use Geocoding to convert those geos into lat/long coordinates. From Google: “Geocoding is the process of converting addresses (like “1600 Amphitheatre Parkway, Mountain View, CA”) into geographic coordinates (like latitude 37.423021 and longitude -122.083739), which you can use to place markers on a map, or position the map.”

To use this Google Maps service, you need to have a Google Maps API key. To get a key, you can follow the directions here. When you sign up for an API key, you should select “Server Side Key,” since we will be running a Python script server-side to access the Google Maps API.

Once you have your api_key, you can work on getting geocoded results for all of your geos. You can do this with the following code:

import requests
# set your google maps api key here.
google_maps_api_key = ''

# get the list of countries from our DataFrame.
countries = grouped_country_df.index
for country in countries:
    # make request to google_maps api and store as json. pass in the geo name to the address 
    # query string parameter.
    url ='{}&amp;key={}'\
         .format(country, google_maps_api_key)
    r = requests.get(url).json()

    # Get lat and long from response. "location" contains the geocoded lat/long value.
    # For normal address lookups, this field is typically the most important.

    lat = r['results'][0]['geometry']['location']['lat']
    lng = r['results'][0]['geometry']['location']['lng']

This only gets you so far, since you still need to do something with those latitude and longitude coordinates. We have a few options here:

  1. If you are building a web application, you can pass those values into an HTML template as variables and they will end up getting plotted via JavaScript.
  2. We can print out the format of the JavaScript, and later past it into our HTML file within script tags.
  3. Other approaches that I’m not going to talk about.

For the sake of time, I’m going to show #2, which lends itself to a one-off analysis. You’d probably want to go with some dynamic templating approach, like #1, if you are going to pull and plot the same data repeatedly.

Add the following code to your for-loop from above, right underneath

lng = r['results'][0]['geometry']['location']['lng']

# set the country weight for later. by getting the value for each index in the dataframe
# as it loops through.
country_weight = int(grouped_country_df.ix[country])
# print out the Javascript that we will be copy-pasting into our HTML file
print '{location: new google.maps.LatLng(%s, %s), weight: %s},' % (lat, lng, country_weight)

After running your script, copy the output, which should look like this:

{location: new google.maps.LatLng(37.09024, -95.712891), weight: 3430},
{location: new google.maps.LatLng(40.463667, -3.74922), weight: 577},
{location: new google.maps.LatLng(55.378051, -3.435973), weight: 352},
{location: new google.maps.LatLng(31.046051, 34.851612), weight: 292},
{location: new google.maps.LatLng(47.516231, 14.550072), weight: 196},
{location: new google.maps.LatLng(-38.416097, -63.616672), weight: 151},
{location: new google.maps.LatLng(20.593684, 78.96288), weight: 133},
{location: new google.maps.LatLng(1.352083, 103.819836), weight: 66},

You’re going to use these values in the next step.

Creating an HTML file that contains Javascript for Plotting your Lat Long Points.

You need to create an HTML file that contains some script tags within it. I am simply going to paste my code below with annotations. If you copy the location strings from above, you will be able to paste them directly into this HTML file under the “heatmapData” array (defined below in the code).

<!DOCTYPE html>
    <title>Simple Map</title>
    <meta name="viewport" content="initial-scale=1.0, user-scalable=no">
    <meta charset="utf-8">
      html, body, #map-canvas {
        height: 100%;
        margin: 0px;
        padding: 0px
    <!-- Load Google Maps API. -->
    function initialize() {
      var heatmapData = [
        {location: new google.maps.LatLng(37.09024, -95.712891), weight: 3430},
        {location: new google.maps.LatLng(40.463667, -3.74922), weight: 577},
        {location: new google.maps.LatLng(55.378051, -3.435973), weight: 352},
        {location: new google.maps.LatLng(31.046051, 34.851612), weight: 292},
        {location: new google.maps.LatLng(47.516231, 14.550072), weight: 196},
        {location: new google.maps.LatLng(-38.416097, -63.616672), weight: 151},
        {location: new google.maps.LatLng(20.593684, 78.96288), weight: 133},
        {location: new google.maps.LatLng(1.352083, 103.819836), weight: 66},
      // Add some custom styles to your google map. This can be a pain. 
      var styles = [ 
          "featureType": "administrative",
          "stylers": [
            { "visibility": "off" }
          "featureType": "road",
          stylers: [ 
            { "visibility": "off"}
          "featureType": "landscape",
          "elementType": "geometry.fill",
          "stylers": [
            { "color": "#ffffff" },
            { "visibility": "on" }
      // create a point on the map for the Atlantic Ocean, 
      // which will later be used for centering the map.
      var atlanticOcean = new google.maps.LatLng(24.7674044, -38.2680446);
      // Create the styled map object.
      var styledMap = new google.maps.StyledMapType(styles, {name:"Styled Map"});
      // create the base map object. put it in the map-canvas id, defined in HTML below.
      map = new google.maps.Map(document.getElementById('map-canvas'), {
        center: atlanticOcean, // set the starting center point as the atlantic ocean
        zoom: 3, // set the starting zoom 
        mapTypeControlOptions: {
          mapTypeIds: [ google.maps.MapTypeId.ROADMAP, 'map_style'] // give the map a type.
      // Create the heatmap object.
      var heatmap = new google.maps.visualization.HeatmapLayer({
        data: heatmapData, // pass in your heatmap data to plot in this layer.
        opacity: 1, 
        dissipating: false, // on zoom, do you want dissipation?
      heatmap.setMap(map); // apply the heatmap to the base map object.
      map.mapTypes.set('map_style', styledMap); // apply the styles to your base map.
      // Add a custom Legend to Your Map
      var legend = document.getElementById('legend');
      // This is hard-coded for the countries I knew existed in the set.
      var country_list = ['USA','Spain','United_Kingdom','Israel',
      // for each country in the country list, append it to the Legend div.

      for (i = 0; i < country_list.length; i++) {
          var div = document.createElement('div');
          div.innerHTML = '<p>' + country_list[i] + '</p>'

     google.maps.event.addDomListener(window, 'load', initialize);


    <'div id="legend" style="background-color:grey;padding:10px;">
    <strong>Countries Mapped</strong>

    <'div id="map-canvas"></div>

Open the HTML file in your browser, and you should see something like this.

google maps heatmap

Et Voila!


Creating and Using SlugFields for URLs with Django

I haven’t written a post centered around coding in a long while, but I’ve recently been learning Django via the Tango with Django tutorial and I got hung up on SlugFields. My problem was that I had never heard of a slugfield, and thus, really had no idea how they would end up being implemented. Slugs? Yes. Fields? Yes. Metal Slug? Yes (great game series)! But SlugFields? No. Nope. Dunno what those are.

SlugField from Django:

Slug is a newspaper term. A slug is a short label for something, containing only letters, numbers, underscores or hyphens. They’re generally used in URLs.”

Related notes from StackOverflow.

Still not getting it? Well, if you are creating a model with a field type Slug, and you use the slugify function, then that field will expect and hold down-cased, space-replaced values. The reason you want a down-cased, space-replaced value is to be able to use that tidy field value in a URL later on. Here is the regex validation that occurs on the slugfield:

slug_re = re.compile(r'^[-a-zA-Z0-9_]+$')
validate_slug = RegexValidator(slug_re, _("Enter a valid 'slug' consisting of letters, numbers, underscores or hyphens."), 'invalid')

In the Tango with Django tutorial, the code for creating your model is given as follows:

from django.template.defaultfilters import slugify
class Category(models.Model):
    name = models.CharField(max_length=128, unique=True)
    views = models.IntegerField(default=0)
    likes = models.IntegerField(default=0)
    slug = models.SlugField()
    def save(self, *args, **kwargs):
        self.slug = slugify(
        super(Category, self).save(*args, **kwargs)

In this case, any time a Category is saved, we are setting the slug attribute to the name of the category by calling the slugify function and passing in

self.slug = slugify(

This is the slugify docstring:

Definition:  slugify(*args, **kwargs)
Converts to lowercase, removes non-word characters (alphanumerics and
underscores) and converts spaces to hyphens. Also strips leading and
trailing whitespace.

After you’ve made the changes to your model and performed your migrations, you can see how this will look when a new category is created in your category table:

mysql> select * from rango_category;
| id | name             | views | likes | slug             |
|  4 | Python           |   128 |    64 | python           |
|  5 | Django           |    64 |    32 | django           |
|  6 | Other Frameworks |    32 |    16 | other-frameworks |
3 rows in set (0.00 sec)

From here, you would use these slugs to create pretty urls by either using the slug itself, or using a combination of the slug/id fields.